seramic: a semi-automatic method for the segmentation of
TRANSCRIPT
HAL Id: hal-03191884https://hal.umontpellier.fr/hal-03191884
Submitted on 8 Apr 2021
HAL is a multi-disciplinary open accessarchive for the deposit and dissemination of sci-entific research documents, whether they are pub-lished or not. The documents may come fromteaching and research institutions in France orabroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, estdestinée au dépôt et à la diffusion de documentsscientifiques de niveau recherche, publiés ou non,émanant des établissements d’enseignement et derecherche français ou étrangers, des laboratoirespublics ou privés.
SEraMic: a semi-automatic method for thesegmentation of grain boundaries
Renaud Podor, X. Le Goff, J. Lautru, H.P. Brau, M. Massonnet, NicolasClavier
To cite this version:Renaud Podor, X. Le Goff, J. Lautru, H.P. Brau, M. Massonnet, et al.. SEraMic: a semi-automaticmethod for the segmentation of grain boundaries. Journal of the European Ceramic Society, Elsevier,In press, �10.1016/j.jeurceramsoc.2021.03.062�. �hal-03191884�
1
SEraMic: a semi-automatic method for the segmentation of grain boundaries
R. Podor*, X. Le Goff, J. Lautru, H.P. Brau, M. Massonnet, N. Clavier
ICSM, Univ Montpellier, CNRS, ENSCM, CEA, Marcoule, France
*Corresponding author
Abstract
The SEraMic method, implemented in the SEraMic plugin for Fiji or ImageJ software, was
developed to calculate a segmented image of a ceramic cross section that shows the grain boundaries.
This method was used to accurately and automatically determine grain boundary positions and further
assess the grain size distribution of monophasic ceramics, metals, and alloys. The only required sample
preparation is polishing the cross section to a mirror-like finish. The SEraMic method is based on at least
six backscattered electron scanning electron microscopy images of a unique region of interest with
various tilt angles ranging from -5° to +5°, which emphasises the orientation contrasts of the grains.
Because the orientation contrast varies with the incident beam angle on the sample, the set of images
contains information related to all the grain boundaries. The SEraMic plugin automatically calculates and
builds a segmented image of the grain boundaries from the set of tilted images. The SEraMic method
was compared with classical thermal etching methods, and it was applied to determine the grain
boundaries in various types of materials (oxides, phosphates, carbides, and alloys). The method remains
easy to use and accurate when the average grain diameter is greater than or equal to 0.25 µm.
Keywords
SEM, BSE, ceramic, segmentation, grain size
2
Introduction
The properties of ceramics are closely related to their microstructure (e.g. grain size, amount and
distribution of porosity, and chemical homogeneity). Thus, measuring the grain size in a ceramic
(ceramography) is an essential step in the characterisation of its microstructure [1]. To assess the grain
size distribution and evaluate the average grain size, it is necessary to collect images with sufficient
contrast at the grain boundaries so that they can be segmented, allowing the determination of the
position of the grain contours. Recording images that contain the information of the grain boundary
position is critical and often difficult to achieve, particularly when the grain size is around or even below
the micrometre scale. It requires the use of techniques that aim to create a contrast between the grains,
which can be difficult and/or time consuming to implement. Two main families of methods are generally
used.
First, grain boundaries can be revealed to create a contrast between the grains. This is achieved
by polishing the surface to the mirror grade and then applying either chemical or thermal
etching. Images of the material surface are further obtained using optical or scanning electron
microscopy. Both methods require either the use of hazardous chemicals [2] or long heat
treatments that often have to be optimised when working with new materials [3]. Finally,
thermal etching could bias the measurements of grain sizes, particularly when grain growth,
surface modifications [4,5,6,7], and/or secondary phase precipitation at the grain boundaries [8]
occur during the heat treatment.
As an alternative, a thin section (approximately 30 µm thick) can be cut from the ceramic. The
images of the grains are then recorded using optical microscopy in the polarised transmitted
light mode. Heilbronner [9] developed an original process, called the Lazy Grain Boundary
method, which is based on sets of polarised micrographs or orientation images. This method
allows fast and automatic/semi-automatic detection of grain boundaries but has only been used
for the study of geological samples using optical microscopy.
Furthermore, methods based on the recording of electron backscatter diffraction (EBSD) maps are
under development [10,11,12]. The primary benefit of grain size measurement using EBSD over imaging
of etched surfaces is that grain boundaries can be precisely identified owing to contrast modification due
to the crystallographic orientation.
3
These different methods are limited by the sample preparation, resolution of the microscope, and/or
availability of an EBSD detector. Finally, these techniques can be very expensive in terms of the time
required for sample preparation.
Once the images are recorded, the next step relies on image processing and segmentation.
Numerous methods currently under development are based on numerical calculations [13,14,15] or
deep learning methods [16,17,18]. Finally, several methods can be used to measure the grain size [19],
including the intercept method [20,21] and standardised methods, such as planimetric grain size
measurement [22]. A set of several hundreds of grains (typically 500–1000) is generally considered to
statistically account for the grain size distribution in the material.
In the present study, we propose the development of a reliable, fast, and easy-to-implement
method for revealing grain boundaries. This method was tested on various ceramics and alloys. An
evaluation criterion of the method is defined to compare the quality of the segmented image obtained
while varying the image processing parameters. The image processing associated with this method was
implemented in the dedicated ‘SEraMic’ plugin for ImageJ [23] and Fiji [24], which allows fast
segmentation of the grain boundaries in ceramics.
Materials and methods
Materials
Five different ceramics and one alloy were used in the present study: (U1-xCex)O2, SnO2, NdPO4, TiC,
YAG, and Ni25Cr. These samples were chosen because they belong to different families of materials, i.e.
oxides, phosphates, carbides, and a Ni25Cr alloy.
Five (U1-xCex)O2 oxides were sintered between 1300 and 1600 °C for 1 to 10 h under an Ar/H2
atmosphere to obtain high-density ceramics with different grain sizes. These ceramics are
electronic conductors and can be directly observed under high-vacuum conditions.
A SnO2:0.5%CoO ceramic was sintered at 1300 °C for 96 h in air [25]. This ceramic is insulating.
This is denoted as SnO2 in the text.
NdPO4 was sintered at 1400 °C for 4 h [26]. This ceramic is insulating.
TiC was sintered at 1700 °C by spark plasma sintering for 5 min at P = 75 MPa [27]. This ceramic
is an electronic conductor.
4
The Ni25Cr (wt.%) alloy was prepared by high-frequency melting from bulk elements with
purities higher than 99.95% [28].
Four yttrium aluminium garnet transparent ceramics (YAG) were provided by IrCer, University of
Limoges (France). They were prepared to have different average grain diameters. These ceramics
are insulating.
Sample preparation
The TiC and SnO2 pellets were embedded in a resin. All the samples were polished to obtain a
mirror-like surface. After a pre-polishing step performed with SiC papers of various grades, a two-step
finishing polish was performed using a 1 µm diamond suspension for 2 min, followed by a 100 nm
colloidal silica suspension for 15 min.
Scanning electron microscopy
An environmental scanning electron microscope (SEM, Quanta 200 ESEM FEG, FEI) was used to
record the images. The internal ring of a directional backscattered electron (BSE) detector (DBS, FEI) was
used to enhance the grain orientation contrast [29]. This detector allows the recording of BSE images at
a low acceleration voltage. It can be used under high-vacuum as well as low-vacuum conditions. The
magnifications of the SEM were calibrated prior to image recording (see Supplementary File S1). With
this calibration step, the measurement errors are limited, and the SEM can be used as an accurate
measurement tool [30].
Description of the SEraMic method
Principle of the SEraMic method
The method is based on the recording of SEM images using the BSE mode. The image contrast
can contain information related to the crystallographic orientation of the grains in the ceramic. These
contrasts are related to the variation in the number of BSEs as a function of the crystallographic
orientation of the grain relative to the incident electron beam. This mode of contrast (also called
channelling contrast) makes it possible to highlight misorientations between grains of identical chemical
5
nature as well as twinning in the grains. These contrasts are at the origin of the eCHORD (electron
CHanneling ORientation Determination) method recently developed by Lafond et al., which aims to
obtain orientation maps of polycrystalline materials using a conventional SEM [31].
When recording channelling contrast images, two neighbouring grains with different
crystallographic orientations can exhibit the same grey level (i.e. the same contrast). Thus, they can be
identified as a single grain in the image. To overcome this bias, it is possible to modify the contrast
between these two grains by changing the incidence angle of the primary electron beam with respect to
the surface of the material. This can be easily achieved by tilting the specimen by a few degrees relative
to the initial sample position [32]. Thus, recording several images at different tilt angles on the same area
of the specimen yields a set of images where all the grains with different crystallographic orientations
can be observed (Figure 1a). Once the images are recorded, the grain boundaries are detected using
image processing (Figure 1b). A schematic representation of the SEraMic method is shown in Figure 1.
Figure 1. Schematic representation of the SEraMic method. a) A series of tilted images is recorded in the
BSE mode. b) The images are processed to extract an average image and the corresponding segmented
image.
6
Optimum SEM operating conditions
The choice of the SEM operating conditions is important to optimise BSE emission and
crystallographic contrast between neighbouring grains and to limit charging effects.
First, the choice of the primary electron beam acceleration voltage (HV) is an important
parameter. When HV ranged between 3 and 10 kV, the contrast could be adjusted to observe the pores
as black zones (grey level equal to 0) without any white pixels. The pores appeared as black and white
zones when HV was higher than 10 kV (see Supplementary File S2). This makes image processing and
pore segmentation more difficult and should be avoided. When HV was lower than or equal to 10 kV, the
grey level associated with the pores did not change with the tilt angle, which further facilitates the
segmentation and recognition of pores in the ceramic. Furthermore, in this range of HV values, the
quantity of emitted BSEs is high (even if it depends on the beam current), and the BSE image resolution
remains below 10–50 nm (even if it strictly depends on the HV value, the average atomic number of the
material, and the working distance) [33]. Thus, high-resolution BSE images can be recorded quite quickly
(within a few seconds to a few tens of seconds). Note that the orientation contrast between two
neighbouring grains also depends on the acceleration voltage of the electron beam. Another possible
method to modulate the orientation contrasts is to record an image series while varying the HV value
(see Supplementary File S3). This possibility has not been exploited in the present study because the
contrast of the pores changes with the HV value, making the segmentation process is less reliable.
Most SEMs can be used in a low-vacuum mode that allows the observation of insulating
materials without any sample preparation (carbon or gold coating). In this mode, where the pressure in
the chamber ranges between 10 and 200 Pa (depending on the apparatus), the BSEs are used to observe
the sample. Primary beam scattering remains low under these gas pressure conditions if the working
distance is minimised (4–8 mm), i.e. the resolution of the BSE images does not decrease [34].
Furthermore, the location of the BSE detector in the SEM chamber is very close to the sample surface
(i.e. the zone where the BSEs are emitted). Consequently, even if they are attenuated by the presence of
gas between the sample and the detector, the quantity of BSEs collected by the detector remains high
(more than 50%) when compared to the quantity of BSEs collected under the same conditions under
high-vacuum conditions. Thus, this imaging mode can be used to observe the grains of an insulating
ceramic without any coating. Carbon coating can also be used, but this will degrade the image contrasts
at very low HV owing to the absorption of low-energy BSEs by the carbon coating.
7
The working distance was set between 6 and 10 mm. As previously explained, a lower value was
set to maximise the collection of BSEs by the detector with respect to the necessity of tilting the sample
and the position of the BSE detector in the SEM. The higher working distance was chosen to work at the
so-called eucentric position (10 mm for the SEM used in this study), which allows the region of interest
(ROI) to remain under the electron beam regardless of the tilt angle chosen. No position adjustment will
be needed to remain at this position, and the recording of the image series will be faster. The range of
working distances depends on the configuration of the SEM that is used and must be adjusted
accordingly.
The upper and lower limits of the tilt angles were +5° and -5°, respectively. For tilt angles outside
this range, the recorded images were slightly deformed, which yields an increase in the size of the grain
boundary during the segmentation process.
Finally, the operating conditions were as follows: The acceleration voltage was set between 6
and 10 kV. The beam current was adjusted to a few hundred picoamperes. When the low-vacuum mode
was used, the gas pressure was adjusted to limit the charging effects while minimising the primary
electron beam skirt effect and BSE absorption by the gas. The gas pressure was generally maintained
between 15 and 22 Pa. All these parameters should be adapted for each type of microscope and/or BSE
detector.
SEraMic image processing
The SEraMic plugin was developed to process the images using Fiji or ImageJ software and extract a
segmented image showing the grain boundaries. The input data consist of a set of tilted images (raw
images or aligned stack of images), and the output data are a stack of aligned images, a binary image
representing grain boundaries and pores, and a composite image showing the position of grain
boundaries and pores on the average image of the stack of aligned images. The workflow follows the
flowchart shown in Figure 2.
The SEraMic plugin can be used as an automatic procedure in which the user provides a set of tilted
images, and the plugin calculates the output images with defined parameters. The plugin can also be
used as a semi-automatic procedure where the user can manually modify the segmented images by
adding pixels (to complete missing grain boundaries) and/or removing pixels (to delete stripes or
supplementary pixels). These sequences are represented by orange rectangles in the flowchart.
8
The flowchart contains three consecutive steps, represented by boxes A, B, and C in Figure 2. Box A
corresponds to the image pre-processing step. A directory (‘analyses’) where the images and results will
be saved is created. The raw images are aligned (‘Linear Stack Alignment with SIFT’ procedure [35]), and
the scale is defined. The ‘aligned stack’ that is computed is saved and will be used during the following
step. Box B corresponds to the core of the plugin where the segmented image representing the positions
of the grain boundaries will be processed. An ROI is defined in the aligned stack. First (sub-box B1), a Z-
projection of the stack allows the calculation of an average intensity image. A threshold of this image
(based on the separation of black pixels) yields a mask representing the position of the pores. Second
(sub-box B2), a series of mathematical operations is applied to the images. The images are first blurred
using a median filter to average the grey level in one grain. Thus, a variance filter is applied to each
image to detect the grey level variation. The resulting images are averaged through a Z-projection using
the maximum intensities of each image. A threshold is then applied to eliminate extra pixels that do not
correspond to the grain boundaries. The resulting image is transformed into a binary image, and several
erode/dilate operations are used to eliminate extra pixels. Then, a mask representing the position of the
grain boundaries is created from this image.
The obtained masks (pores and grain boundaries) can be manually adjusted by adding or removing
pixels if a semi-automatic procedure has been chosen. A composite image is created using the pore
mask, grain-boundary mask, and average image, which is then saved. A mask image, which is a
segmented image showing the positions of the pores and grain boundaries, is saved. Last, box C is an
optional operation that allows the determination of the grain sizes from the mask image. The results are
saved in a ROI.zip file.
The SEraMic plugin was developed in Java and can be downloaded from the GitHub repository
(https://github.com/xavier-legoff/SEraMic) for research purposes. It was designed to work with Fiji
(minimum version required is ImageJ 1.53 g) and ImageJ (with the ‘Linear Stack Alignment with SIFT’
plugin installed).
9
10
Figure 2. Flowchart of the SEraMic plugin for grain-boundary segmentation. (a) Image pre-processing, (b)
Image processing and grain-boundary segmentation, (c) Calculation of grain sizes.
Results and discussion
The method and plugin were tested on five different materials. The objectives are, on the one hand,
to determine the reliability of the segmentation and to evaluate the limits of the method, and on the
other, to demonstrate the effectiveness of the method on different chemical families of materials.
Types of materials that can be studied
The SEraMic method has been used to record tilted image series and to determine the segmented
images of the chosen materials. Segmented images were obtained successfully on each sample from a
series containing 6–10 images. An example of a segmented image for each material is shown in Figure 3.
These examples illustrate the possibility of using the SEraMic method for very different materials and
under various conditions (low/high vacuum, carbon-coated/uncoated sample) as well at different
magnifications. Images of the NdPO4, TiC, and SnO2 samples (Figure 3a–c) were recorded under low-
vacuum conditions so that the charging effect due to their insulating nature was well balanced by the
presence of gas in the SEM chamber. The segmented images obtained (illustrated by the red lines in the
images) were located exactly at the grain boundaries. In Figure 3d–f, the images were recorded under
high-vacuum conditions because the studied samples were electronic conductors (Ni25Cr alloy and
(U0.9Ce0.1)O2 ceramic) or were carbon-coated to eliminate surface charges due to the electron beam
(SnO2). Once again, the grain boundaries were successfully identified using the SEraMic method
(illustrated in red in the corresponding images). It should be noted that the potential negative effect of
the carbon coating on the SEM images (i.e. attenuation of the BSE signal at low acceleration voltages)
remained limited, and images were recorded at an acceleration voltage of as low as 3 kV (see
Supplementary File S2 and S3).
11
Figure 3. Segmented images obtained for various materials. a) Uncoated NdPO4 observed under low-
vacuum conditions. b) Uncoated TiC observed under low-vacuum conditions. c) Uncoated SnO2 observed
under low-vacuum conditions. d) Carbon-coated SnO2 observed under high-vacuum conditions. e)
Uncoated Ni25Cr alloy observed under high-vacuum conditions. f) Uncoated (U0.9Ce0.1)O2 observed under
high-vacuum conditions.
Definition of accuracy criteria
Accuracy criteria were defined to qualify segmentation quality. First, an automatically segmented image
of the ceramic was obtained using the SEraMic plugin. This segmented image was manually corrected to
complete missing grain boundaries and delete pixels that did not correspond to pores or grain
boundaries, i.e. the segmented image was cleaned of all biases induced by the pores and the missing
grain boundaries to produce a ‘best estimate’ image. This image is called the reference segmented
12
image. Then, the reference segmented image was compared with the automatically segmented image.
The number of pixels that were added to complete the automatically segmented image, as well as the
number of pixels that were erased, were used as measurement criteria of fidelity and accuracy
(hereafter called accuracy criteria). These criteria are expressed as the percentage of pixels added (or
removed) relative to the total number of pixels in the reference segmented image. Thus, higher values of
these criteria indicate lower quality of the automatic segmentation. The criteria are denoted as ‘AC%+’
and ‘AC%-’, respectively, in the remainder of the text. They will be used to compare the various
automatically segmented images and determine the quality of the automatically segmented images.
Minimum number of tilted images to be recorded
The minimum number of images required to obtain good quality segmentation was determined for
the different materials studied. A series of ten tilted images were recorded (Figure 4a), and reference
segmented images were generated (Figure 4b and c). Then, different series of n images (1 ≤ n ≤ 10) were
used to generate automatically segmented images of the grain boundaries, i.e. without manual
correction. The as-obtained segmented images contained missing grain boundaries and extra pixels
corresponding to residual polishing scratches (see Figure 4d and e). The segmented images were
compared with the reference segmented image to determine the values of the AC%+ and AC%- accuracy
criteria. The values are reported in Figure 5 as a function of the number of images, n, used to calculate
each segmented image.
13
Figure 4. a) Series of ten BSE images of a (U0.9Ce0.1)O2 ceramic recorded with various tilt angles. b)
Average image obtained from the ten images. c) Segmented image obtained from the ten images using
the SEraMic plugin with correction (used as the reference segmented image) d) Average image obtained
using the two images recorded with -1° and 0° tilt angles. e) Segmented image obtained without
correction using the SEraMic plugin with images recorded at 5° and 6°. Scale bars = 5 µm.
The data reported in Figure 5a indicate that the number of pixels removed (AC%- criterion) was
independent of the number of images considered for image segmentation. Most of these pixels
correspond to polishing defects (mainly scratches) and are analysed by the automatic segmentation
procedure as grain boundaries, independent of the number of tilted images used for segmentation. This
clearly indicates that the quality of the polishing is one of the main parameters that limits the efficiency
of the automatic recognition of grain boundaries.
The number of pixels added to the segmented images (AC%+ criterion) decreased as a function of
the number of images (n) used for the calculation of the segmented image. When this criterion is lower
than approximately 0.5%, this corresponds to the addition of a few (typically 1–5) grain boundaries on
the images that are not automatically detected (see Supplementary File S4). In this case, the number of
missing grain boundaries is sufficiently low to not affect the calculation of the average grain size, and this
14
value for the AC%+ criterion can be an acceptable limit for the calculation of a high-quality segmented
image. When considering the different materials studied, 6–10 images are generally suitable for
achieving this AC%+ value, i.e. to automatically determine a high-quality segmented image without
further manual correction. It appears that this range of values is independent of the magnification and
the material studied.
Figure 5. Variation of the (a) AC%- and (b) AC%+ values (expressed in percentage) for segmented images
obtained with n tilted images as a function of n. These parameters were measured for (U0.9Ce0.1)O2 at
two different magnifications; TiC, SnO2, and NdPO4 ceramics; and a Ni25Cr alloy.
15
Determination of the tilt angles
The choice of tilt angles can first appear as an important parameter for image segmentation
processing. Thus, ten different sets of four out of the ten tilted images (previously recorded on a
(U0.9Ce0.1)O2 ceramic) were used to extract the grain boundaries and determine the number of grains as
well as their average size. The average number of grains that was determined for every set of four
images was 368 (± 8), while the number of grains determined from the complete dataset of ten images
was 366 (Figure 6a). Thus, most of the grains present in the ROI were detected independently of the
chosen tilt angles.
In parallel, the average grain diameter was measured from each set of four random images. The
ten values are shown in Figure 6a. From these ten values, the mean value of the average grain diameters
measured from the four-image random series was determined. It was equal to 2.34 (± 0.01) µm, while
the average grain diameter determined for the entire ten tilted image series was 2.2 µm (Figure 6b).
Thus, these values are very close, and the choice of tilt angles did not significantly modify the average
grain diameter. The AC%+ and AC%- criteria values associated with each four-image random series are
reported in Figure 6b as a function of the randomly selected images. Once again, the values of the
accuracy criteria did not vary significantly according to the series considered.
These results indicate that the quality of the automatically segmented image does not depend
on the tilt angles between the images. Thus, the tilt angles used to record the tilted image series can be
chosen randomly between -5° and +5° without selecting specific tilt angles.
16
Figure 6. a) Average grain diameter and number of grains associated with the ten series of four images
randomly selected within the dataset of ten images. b) AC%+ and AC%- values associated with the ten
series of four images randomly selected within the dataset of ten images.
Recording 6–10 images with close but different tilt angles is a good compromise to obtain highly
contrasted grain boundaries and to automatically and accurately determine a segmented image of the
grains that are present in the ROI. The time required to record the SEM tilted image series was generally
less than 15 min, while that required to compute the segmented image was less than 5 min.
Dependence on the user
The quality of image segmentation, which is often performed semi-manually or fully manually,
can be user dependent. As some manual corrections can be performed using the SEraMic plugin by
adding or deleting a few points, the average grain sizes determined from the same set of images by three
different users are presented in Table 1 for comparison. The three sets of data were equivalent, and the
result obtained, in terms of average grain size, was independent of the user.
17
Table 1. Compositions, sintering temperatures, and sintering times of the (U1-xCex)O2 ceramics. The
average grain diameters (D) and associated standard deviations (σD) that were determined in the present
study using the SEraMic method are also reported.
User 1 User 2 User 3
x T (°C) Time
(h)
Atm D (µm) σD (µm) D (µm) σD (µm) D (µm) σD (µm)
0.25 1300 10 Ar 2.06 1.05 2.15 0.95 1.98 1.10
0.25 1300 10 Ar-H2 0.14 0.08 0.15 0.07 0.17 0.09
0.14 1600 10 Ar-H2 7.01 4.05 7.30 3.48 7.76 3.73
0.10 1500 5 Ar-H2 1.51 0.71 1.58 0.55 1.49 0.68
0.10 1500 1 Ar-H2 0.75 0.27 0.68 0.29 0.67 0.34
0.10 1300 10 Ar-H2 0.18 0.08 0.14 0.07 0.20 0.08
Minimum grain size
Another important parameter is the minimum size of the average grain distribution that can be
determined using the SEraMic method. To determine this, (U1-xCex)O2 ceramics with various average
grain sizes were studied. The average grain diameters (and associated standard deviations) are listed in
Table 1, and the accuracy criteria associated with each ceramic were calculated. For the (U1-xCex)O2
ceramics containing the smallest grains (average grain size of approximately 150 nm), the images were
recorded using magnifications higher than 20 k (corresponding to a 3.186 × 2.751 µm² image for 1024 ×
884 pixels). Under these conditions, the vacuum quality in the SEM chamber was not sufficient to record
more than six tilted images. Indeed, the sample surface was rapidly contaminated by the residual
molecules present in the SEM chamber. Thus, the orientation contrasts were attenuated. This effect
strongly limited the image quality, and thus the quality and reliability of the segmented image
calculation.
The AC%+ (number of pixels added to the segmented images to complete the position of the
grain boundaries) and AC%- criteria rapidly decreased when the average grain size increased (Figure 7).
Considering that the SEraMic method allows the accurate and automatic determination of the grain
boundary position when the AC%+ criteria is lower than 0.5, the lower limit of the grain size (diameter)
that could be determined by the SEraMic method was close to 0.25 µm with the configuration of the
18
microscope used in this study. Apart from nanostructured materials, this allows the study of the majority
of ceramic materials that are classically studied.
Figure 7. Variation in the AC%+ and AC%- values as a function of the average grain size for the (U1-xCex)O2
compounds. Corresponding averaged and segmented images are reported in Supplementary File S5.
Comparison of SEraMic with thermal etching
The robustness of the SEraMic method was tested by comparing the segmentation and average grain
diameters determined using the SEraMic and thermal etching methods. To achieve this comparison, a
series of transparent YAG ceramics was prepared by sintering at various temperatures and dwell times to
monitor the average grain size. The image recording and segmentation processes were achieved after
either thermal etching or recording of tilted image series on a mirror-like surface using the SEraMic
method. The average grain sizes were determined automatically using the same procedure from the
different segmented images (Table 2). On YAG3 and YAG4, thermal etching was performed first, and
several zones of the samples were characterised (TEtch_1 in Table 2). The samples were then polished to
obtain a mirror-like surface, and the same zones were characterised again using the SEraMic procedure
(SEraMic in Table 2). Finally, all the samples were thermally etched, and the average grain size was
characterised in the same zones (TEtch_2 in Table 2). A series of SEM images and related segmented
images recorded on YAG4 are shown in Figure 8. The same grains were detected regardless of the
19
segmentation method considered (see Supplementary File S6). The grain sizes were determined from the
grains observed in the segmented images. A total of 212 to 229 grains were observed in these images.
The grain size distributions were very similar and independent of the segmentation method. The average
grain sizes that were determined from each segmented image were 1.31 ± 0.83 µm (after the first
thermal etching), 1.28 ± 0.81 µm (after sample polishing and use of the SEraMic method), and 1.27 ±
0.79 µm (after the second thermal etching). The variation in the average grain size remained limited and
within the statistical tolerance interval, regardless of the segmentation method.
Table 2. Average grain diameters and associated standard deviations (expressed in µm) obtained from
the YAG sample series of pellets determined from segmented images obtained using various techniques
(TEtch = thermal etching). The average grain diameters determined from the same ROI by different
methods are reported on the same line. Each line corresponds to analyses performed on different ROIs.
(ND = not determined).
YAG1 YAG2 YAG3 YAG4 SEraMic TEtch_2 SEraMic TEtch_2 TEtch_1 SEraMic TEtch_2 TEtch_1 SEraMic TEtch_2 0.97 ± 0.41
0.96 ± 0.42
8.82 ± 4.70
ND 3.18 ± 1.67
3.24 ± 1.67
3.16 ± 1.58
1.29 ± 0.81
1.23 ± 0.83
1.24 ± 0.81
0.95 ± 0.48
0.94 ± 0.47
9.24 ± 4.84
ND 3.00 ± 1.47
2.94 ± 1.48
2.94 ± 1.49
1.27 ± 0.70
1.24 ± 0.70
1.23 ± 0.70
1.01 ± 0.56
1.05 ± 0.51
9.87 ± 4.58
ND 2.99 ± 1.72
3.07 ± 1.70
3.04 ± 1.70
1.25 ± 0.79
1.26 ± 0.80
1.20 ± 0.80
0.99 ± 0.47
1.04 ± 0.45
9.55 ± 5.08
ND 3.09 ± 1.56
3.14 ± 1.56
3.10 ± 1.62
1.31 ± 0.83
1.28 ± 0.81
1.27 ± 0.79
9.01 ± 4.88
ND ND 2.72 ± 1.64
2.74 ± 1.67
1.34 ± 0.79
1.32 ± 0.80
1.29 ± 0.80
7.72 ± 3.68
7.65 ± 3.81
ND 2.76 ± 1.67
2.72 ± 1.70
7.91 ± 3.545
8.06 ± 3.46
ND 2.64 ± 1.57
2.63 ± 1.57
7.46 ± 3.93
7.51 ± 3.80
ND 40.83 ± 20.46
42.60 ± 20.35
7.83 ± 3.86
8.11 ± 4.10
ND 44.08 ± 24.63
42.60 ± 20.35
20
Raw image Segmented image Grain size distribution (µm) Th
erm
al e
tchi
ng 1
SEra
Mic
Ther
mal
etc
hing
2
Figure 8. Images recorded after the first thermal etching, after polishing for SEraMic analysis, and after
the second thermal etching. All images were recorded on the same zone of the YAG4 sample.
Corresponding segmented images and grain size distributions are also reported (approximately 210–220
grains were considered in these images).
The average grain sizes determined using the SEraMic procedure are plotted as a function of the
average grain size values obtained after thermal etching for comparison in Figure 9. There is a perfect
match between the two datasets, indicating that the SEraMic method yields the same average grain
values as the conventional method (image segmentation from images recorded on thermally etched
materials). This is true for a large range of grain diameters (1–44 µm for the series of materials studied
herein). Furthermore, the standard deviation around the mean grain size was identical, independent of
the segmentation method. This is due to the fact that the same grains were recognised, and this
indicates that the SEraMic method leads to a grain size dataset that is identical to the one provided by
21
classical methods. The SEraMic segmentation method is a robust method that combines reproducibility
and reliability.
Figure 9. Comparison of the average grain sizes determined using a segmented image obtained with the
SEraMic plugin on polished samples and after thermal etching on YAG samples. The same zones of the
different samples were systematically studied.
In some cases, thermal etching can lead to precipitation or exudation of secondary phases. This
phenomenon was observed locally in the YAG2 sample (Supplementary File S7). This can modify the
sample microstructure and degrade the properties of the sample. Thus, thermal etching should be
avoided with this type of material. In this case, the SEraMic method is particularly well adapted because
no thermal treatment is required for image segmentation processing.
Conclusion
The SEraMic method, implemented in the SEraMic plugin for Fiji and ImageJ software, was
developed to calculate a segmented image from a ceramic cross section that contains the information
related to the grain boundaries. This method can be suitably used to determine the grain boundary
positions of monophasic ceramics, metals, and alloys.
22
The SEraMic method is particularly well adapted for the characterisation of ceramic materials. It is
fast, easy to implement, and provides very accurate segmented images showing grain boundaries and
pores. The preparation of the cross section is one of the key factors for the fast calculation of an accurate
segmented image because the sample surface must be polished to a mirror-like surface. The quality of
the polishing should be perfect to eliminate all remaining scratches at the surface of the sample.
Recording the images only requires an SEM equipped with a BSE detector operating at a low acceleration
voltage (3–10 kV), and a sample holder that can be tilted between -5° and +5°. Moreover, it is important
to note that coating the sample is generally not required because the low-vacuum mode implemented in
most conventional SEMs is suitable for recording BSE images on insulating materials. However, carbon
coating remains a valid option because it does not modify the efficiency of the SEraMic method, but it
does require a higher acceleration voltage (typically 8–10 kV).
With these few constraints, the SEraMic method provides an efficient way to accurately characterise
the grain distribution of conductive and insulating materials with average grain diameters of as low as
250 nm. The acquisition time of the SEM tilted image series, which does not require any additional
sample preparation after the polishing step, is low and typically ranges between 3 and 15 min for a series
of six to ten images. The total time between the beginning of image recording and obtaining the final
segmented image was less than 10 min. Moreover, the segmentation step is fully automated, with the
possibility of modifying some parameters or correcting the automatically generated segmented image.
Thus, all the grains present in the region of interest are detected automatically. In addition, because the
procedure can be fully automated, it is important to underline that the segmented image does not
depend on the choices of the operator.
The SEraMic method (and the SEraMic plugin) was implemented to study various types of
ceramics (oxides, phosphates, and carbides) with results that are in good correlation with the grain sizes
obtained using classical methods. It allows the researcher to save time by providing data related to the
sample microstructure very quickly. The method can probably be extended to biphasic ceramics, but no
trial has been performed during the development of the SEraMic method.
Acknowledgments: The authors would like to thank Yann Launay and Camille Perrière (IrCer, Limoges,
France) for providing the YAG samples.
23
References
1 M.I. Mendelson, Average grain size in ceramics, J. Am. Ceram. Soc. 52 (1969) 443-446. https://doi.org/10.1111/j.1151-2916.1969.tb11975.x 2 J.D. Katz, G. Hurley, Etching alumina with molten vanadium pentoxide, J. Am. Ceram. Soc. 73 (1990) 2151-2152. https://doi.org/10.1111/j.1151-2916.1990.tb05292.x 3 U. Täffner, V. Carle, U. Schäfer, M.J. Hoffmann, (2004). Preparation and microstructural analysis of high-performance ceramics, in: G.F. Vander Voort (Ed.), Metallography and Microstructures, Materials Park, OH: ASM International, 2004, pp. 1057-1066. https://doi.org/10.31399/asm.hb.v09.a0003795 4 I.O. Owate, R. Freer, Thermochemical etching method for ceramics, J. Am. Ceram Soc. 15 (1992) 1266-1268. https://doi.org/10.1111/j.1151-2916.1992.tb05567.x 5 C. Garcıa de Andrés, F.G. Caballero, C. Capdevila, D. San Martın, Revealing austenite grain boundaries by thermal etching: advantages and disadvantages, Mater. Charact. 49 (2002) 121-127. https://doi.org/10.1016/S1044-5803(03)00002-0 6 Y. Palizdar, D. San Martin, M. Ward, R.C. Cochrane, R. Brydson, A.J. Scott, Observation of thermally etched grain boundaries with the FIB/TEM technique, Mater. Charact. 84 (2013) 28-33. https://doi.org/10.1016/j.matchar.2013.07.003 7 T. Spusta, M. Jemelka, K. Maca, The comprehensive study of the thermal etching conditions for partially and fully dense ceramic samples, Sci. Sinter., 51 (2019) 257-264. https://doi.org/10.2298/SOS1903257S 8 A.A. Felix, V.D.N. Bezzon, M.O. Orlandi, D. Vengust, M. Spreitzer, E. Longo, D. Suvorov, J.A. Varela, Role of oxygen on the phase stability and microstructure evolution of CaCu3Ti4O12 ceramics, J. Eur. Ceram. Soc. 37 (2017) 129–136. https://doi.org/10.1016/j.jeurceramsoc.2016.07.039 9 R. Heilbronner, Automatic grain boundary detection and grain size analysis using polarization micrographs or orientation images, J. Struct. Geol. 22 (2000) 969-981. https://doi.org/10.1016/S0191-8141(00)00014-6 10 A. Coutinho, S.C.K. Rooney, E.J. Payton, Analysis of EBSD grain size measurements using microstructure simulations and a customizable pattern matching library for grain perimeter estimation, Metall. Mater. Trans. A 48A (2017) 2375-2395. https://doi.org/10.1007/s11661-017-4031-z 11 H. Chen, Y. Yao, J. A. Warner, J. Qu, F. Yun, Z. Ye, S. P. Ringer, R. Zheng, Grain size quantification by optical microscopy, electron backscatter diffraction, and magnetic force microscopy, Micron 101 (2017) 41–47. http://dx.doi.org/10.1016/j.micron.2017.06.001 12 M. Ben Saada, X. Iltis, N. Gey, B. Beausir, A. Miard, P. Garcia, N. Maloufi, Influence of strain conditions on the grain sub-structuration in crept uranium dioxide pellets, J. Nucl. Mater. 518 (2019) 265-273. https://doi.org/10.1016/j.jnucmat.2019.02.052 13 R. Chinn, Grain sizes of ceramics by automated image analysis, J. Am. Ceram. Soc. 77 (1994) 589-592. https://doi.org/10.1111/j.1151-2916.1994.tb07033.x 14 S. Wang, J. Waggoner, J. Simmons, Research summary graph-cut methods for grain boundary segmentation. JOM 63 (2011) 49-51. https://doi.org/10.1007/s11837-011-0111-5 15 A. Campbell, P. Murray, E. Yakushina, S. Marshall, W. Ion, New methods for automatic quantification of microstructural features using digital image processing, Mater. Des. 141 (2018) 395-406. https://doi.org/10.1016/j.matdes.2017.12.049 16 O. Dengiz, A.E. Smith, I. Nettleship, Grain boundary detection in microstructure images using computational intelligence, Computers in Industry, 56 (2005) 854-866. https://doi.org/10.1016/j.compind.2005.05.012 17 M. Jungmann, H. Pape, P. Wißkirchen, C. Clauser, T. Berlage, Segmentation of thin section images for grain size analysis using region competition and edge-weighted region merging, Computers & Geosciences, 72 (2014) 33-48. https://doi.org/10.1016/j.cageo.2014.07.002 18 M. Li, D. Chen, S. Liu, F. Liu, 2020. Grain boundary detection and second phase segmentation based on multi-task learning and generative adversarial network, Measurement 162, 107857. https://doi.org/10.1016/j.measurement.2020.107857 19 G.F. Vander Voort, Grain size measurement methods: A review and comparison, Microsc. Microanal. 19 (2013) 1760-1761. https://doi:10.1017/S1431927613010799
24
20 H. Abrams, Grain size measurements by the intercept method, Metallography, 4 (1971) 59-78. https://doi.org/10.1016/0026-0800(71)90005-X 21 H. Abrams, Practical applications of quantitative metallography," in Stereology and Quantitative Metallography, ed. G. Pellissier and S. Purdy (West Conshohocken, PA: ASTM International, 1972), 138-182. https://doi.org/10.1520/STP36848S 22 G.F. Vander Voort, Committee E-4 and Grain Size Measurements: 75 Years of Progress, Standardization News 19, (1991) 42-47. https://www.metallography.com/grain.htm 23 C.A. Schneider, W.S. Rasband, K.W. Eliceiri, NIH Image to ImageJ: 25 years of image analysis, Nat. Methods 9 (2012) 671-675. https://doi.org/10.1038/nmeth.2089 24 J. Schindelin, I. Arganda-Carreras, E. Frise, V. Kaynig, M. Longair, T. Pietzsch, S. Preibisch, C. Rueden, S. Saalfeld, B. Schmid, J.-Y. Tinevez, D.J. White, V. Hartenstein, Kevin Eliceiri, P. Tomancak, A. Cardona, Fiji: an open-source platform for biological-image analysis, Nat. Methods 9 (2012) 676-682. https://doi.org/10.1038/nmeth.2019 25 A. Maître, D. Beyssen, R. Podor, Modelling of the grain growth and the densification of SnO2-based ceramics, Ceram. Int. 34 (2008) 27-35. https://doi.org/10.1016/j.ceramint.2006.07.008 26 D. Qin, A. Mesbah, J. Lautru, S. Szenknect, N. Dacheux, N. Clavier, Reaction sintering of rhabdophane into monazite-cheralite Nd1-2xThxCaxPO4 (x = 0 – 0.1) ceramics, J. Eur. Ceram. Soc. 40 (2020) 911-922. https://doi.org/10.1016/j.jeurceramsoc.2019.10.050 27 H. Aréna, M. Coulibaly, A. Soum-Glaude, A. Jonchère, A. Mesbah, G. Arrachart, N. Pradeilles, M. Vandenhende, A. Maitre, X. Deschanels, SiC-TiC nanocomposite for bulk solar absorbers applications: Effect of density and surface roughness on the optical properties, Sol. Energy Mater. Sol. Cells 191 (2019) 199-208. https://doi.org/10.1016/j.solmat.2018.11.018 28 T. Perez, L. Latu-Romain, R. Podor, J. Lautru, Y. Parsa, S. Mathieu, M. Vilasi, Y. Wouters, In situ oxide growth characterization of Mn-containing Ni–25Cr (wt%) model alloys at 1050°C, Oxid. Met. 89 (2018) 781–795. https://doi.org/10.1007/s11085-017-9819-0 29 Application Note: Information from Every Angle. Directional BSE Detector for Next-Level Imaging, http://emc.missouri.edu/wp-content/uploads/2016/03/FEI-Information-for-Every-Angle-DBS-Detector.pdf 30 D. B. Ballard, A procedure for calibrating the magnification of a scanning electron microscope using NBS SRM 484, NBSIR 77-1248, U.S. Department of Commerce, National Bureau of Standards, June 1977. https://nvlpubs.nist.gov/nistpubs/Legacy/IR/nbsir77-1248.pdf 31 C. Lafond, T. Douillard, S. Cazottes, P. Steyer, C. Langlois, Electron CHanneling ORientation Determination (eCHORD): An original approach to crystalline orientation mapping. Ultramicroscopy 186 (2018) 146-149. https://doi.org/10.1016/j.ultramic.2017.12.019 32 D. Gruner, Z. Shen, Direct scanning electron microscopy imaging of ferroelectric domains after ion milling, J. Am. Ceram. Soc. 93 (2010) 48–50. https://doi.org/10.1111/j.1551-2916.2009.03392.x 33 R. Wuhrer, K. Moran, Low voltage imaging and X-ray microanalysis in the SEM: challenges and opportunities, IOP Conf. Ser.: Mater. Sci. Eng. 109 (2016) 012019. https://doi.org/10.1088/1757-899X/109/1/012019 34 D.J. Stokes, Principles and practice of variable pressure/environmental scanning electron microscopy (VP-ESEM). 2008 John Wiley & Sons, Ltd. 221 pages. https://doi.org/10.1002/9780470758731 35 D. Lowe, Distinctive Image Features from Scale-Invariant Keypoints. Int. J. Comput. Vis. 60 (2004) 91-110. https://doi.org/10.1023/B:VISI.0000029664.99615.94